Dual‐Wavelength Synaptic Simulator ReS₂/Ta₂NiSe₅ for Multi‐Timescale Learning in Neuromorphic Computing

Abstract To address the limitations of silicon‐based devices in neuromorphic computing, this study proposes a dual‐wavelength photomodulated synaptic device based on the ReS₂/Ta₂NiSe₅ heterojunction. Through heterojunction band engineering, wavelength‐selective synaptic plasticity is achieved, leveraging the photothermal effect at 1550 nm wavelength and the photoelectric effect at 520 nm wavelength. This enables a wavelength‐selective synaptic weight update strategy, addressing the limitation of conventional single‐wavelength devices that lack multi‐timescale adaptability. The device exhibits distinct relaxation timescales, with slow synaptic weight updates (≈17 s) at 1550 nm wavelength due to photothermal‐driven charge trapping, and (≈6 s) at 520 nm wavelength enabled by efficient photoelectric carrier dynamics. This dual‐timescale synaptic plasticity enables precise control over synaptic weight adaptation, supporting multi‐timescale learning. Validation through LeNet convolutional neural network (CNN) training on the Extended MNIST (EMNIST) dataset confirms that the dual‐wavelength cooperative learning strategy significantly improves both training efficiency and generalization, achieving a final accuracy of 95.5%. These findings highlight the potential of dual‐wavelength heterojunction photonic synaptic devices for adaptive neuromorphic computing, offering a scalable approach for multi‐timescale learning in future photonic artificial intelligence (AI) architectures.